minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

use envs in server_args (#22994)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-04-16 15:01:33 -07:00
committed by GitHub
parent c0172aef6e
commit c83ef4fdb6
2 changed files with 16 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/environ.py
+2
View File
@@ -315,8 +315,10 @@ class Envs:
# AMD & ROCm
SGLANG_USE_AITER = EnvBool(False)
SGLANG_USE_AITER_UNIFIED_ATTN = EnvBool(False)
SGLANG_ROCM_FUSED_DECODE_MLA = EnvBool(False)
SGLANG_ROCM_DISABLE_LINEARQUANT = EnvBool(False)
SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK = EnvInt(4096)
# MPS (Apple Silicon)
SGLANG_USE_MLX = EnvBool(False)
python/sglang/srt/server_args.py
+14 -16
View File
@@ -37,12 +37,10 @@ from sglang.srt.utils.common import (
LORA_TARGET_ALL_MODULES,
SUPPORTED_LORA_TARGET_MODULES,
cpu_has_amx_support,
get_bool_env_var,
get_device,
get_device_memory_capacity,
get_device_name,
get_device_sm,
get_int_env_var,
get_nvidia_driver_version,
get_quantization_config,
human_readable_int,
@@ -1025,7 +1023,7 @@ class ServerArgs:
self.mm_process_config = {}
# Handle ModelScope model downloads
if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
if envs.SGLANG_USE_MODELSCOPE.get():
self._handle_modelscope_paths()
# Mamba scheduler strategy
@@ -1851,13 +1849,13 @@ class ServerArgs:
"Detected SM120 and MXFP4 quantization format for GPT-OSS model, enabling triton_kernel MOE kernel."
)
elif (
is_hip() and get_bool_env_var("SGLANG_USE_AITER")
is_hip() and envs.SGLANG_USE_AITER.get()
) and is_mxfp4_quant_format:
self.moe_runner_backend = "auto"
logger.warning(
"Detected ROCm and MXFP4 quantization format for GPT-OSS model, enabling aiter MXFP4 MOE kernel."
)
elif is_hip() and get_bool_env_var("SGLANG_USE_AITER"):
elif is_hip() and envs.SGLANG_USE_AITER.get():
# For GPT-OSS bf16 on ROCm with aiter, use triton backend
# because aiter CK kernel doesn't support all GEMM dimensions
self.moe_runner_backend = "triton"
@@ -2175,7 +2173,7 @@ class ServerArgs:
if (
model_arch in ["Qwen3VLForConditionalGeneration"]
and is_hip()
and get_bool_env_var("SGLANG_USE_AITER_UNIFIED_ATTN")
and envs.SGLANG_USE_AITER_UNIFIED_ATTN.get()
and self.page_size is None
):
self.page_size = 16
@@ -2851,7 +2849,7 @@ class ServerArgs:
"FlashInfer TRTLLM routed MoE is enabled. --disable-shared-experts-fusion is automatically set."
)
if get_bool_env_var("SGLANG_CUTLASS_MOE"):
if envs.SGLANG_CUTLASS_MOE.get():
logger.warning(
"SGLANG_CUTLASS_MOE is deprecated, use --moe-runner-backend=cutlass and/or --speculative-moe-runner-backend=cutlass instead"
)
@@ -2911,12 +2909,12 @@ class ServerArgs:
logger.warning(
f"Ascend fused EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
)
fuse_mode = os.environ.get("SGLANG_NPU_FUSED_MOE_MODE", None)
if fuse_mode not in ["1", "2"]:
fuse_mode = envs.SGLANG_NPU_FUSED_MOE_MODE.get()
if fuse_mode not in [1, 2]:
raise ValueError(
f"Wrong value of {fuse_mode=}, the NPU only support 1 or 2."
)
elif fuse_mode == "2":
elif fuse_mode == 2:
assert (
self.quantization == "modelslim"
), "When fuse_mode is set to 2, the NPU supports only ModelSlim quantization."
@@ -2931,7 +2929,7 @@ class ServerArgs:
)
if self.deepep_mode != "auto":
logger.warning("--deepep-mode is ignored for Flashinfer MoE A2A")
if os.environ.get("SGLANG_MOE_NVFP4_DISPATCH") is None:
if not envs.SGLANG_MOE_NVFP4_DISPATCH.is_set():
envs.SGLANG_MOE_NVFP4_DISPATCH.set(True)
logger.warning(
"SGLANG_MOE_NVFP4_DISPATCH is set to True for Flashinfer MoE A2A"
@@ -2953,9 +2951,9 @@ class ServerArgs:
# Skip validation if chunked prefill is disabled (i.e., size <= 0).
# Skip validation if disaggregation mode is decode.
if self.chunked_prefill_size > 0 and self.disaggregation_mode != "decode":
assert (self.chunked_prefill_size) <= get_int_env_var(
"SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK", 4096
), "SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK (default 4096) must be larger or equal to chunked_prefill_size"
assert (
self.chunked_prefill_size
) <= envs.SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK.get(), "SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK (default 4096) must be larger or equal to chunked_prefill_size"
def _handle_eplb_and_dispatch(self):
if self.enable_eplb and (self.expert_distribution_recorder_mode is None):
@@ -3739,9 +3737,9 @@ class ServerArgs:
self.enable_deterministic_inference = True
# For VLM
os.environ["SGLANG_VLM_CACHE_SIZE_MB"] = "0"
envs.SGLANG_VLM_CACHE_SIZE_MB.set(0)
# TODO remove this environment variable as a whole
os.environ["SGLANG_ENABLE_DETERMINISTIC_INFERENCE"] = "1"
envs.SGLANG_ENABLE_DETERMINISTIC_INFERENCE.set(True)
if self.enable_deterministic_inference:
if self.enable_aiter_allreduce_fusion: